• Variables irrelevant to analysis are removed.
• Missing/Incorrect Values:
  • Missing zip code values filled using the uszipcode package.
    • Unspecified boroughs imputed using the zip code.
  • Rows with missing ‘SCORE’ and ‘GRADE’ values are removed.
  • Incorrect ‘VIOLATION CODE’ values are removed.

• Joined restaurant inspection data with violation code data from the Violation Health Code Mapping dataset.

  • Includes important categorical information about each violation.

• Created new variables:

  • ‘VIOLATION COUNT’: Number of violations for each inspection.
  • ‘CRITICAL COUNT’: Number of critical violations for each inspection. A critical violation is one that could cause foodborne illness.
  • ‘DAY TYPE’: Whether the inspection was on a weekday or weekend.
  • 23 new variables for each violation category.
    • Each variable represents the number of violations in each category for each inspection.

• Data is now grouped by restaurant and inspection date.

  • Each row represents a single inspection for a restaurant.

• The Kruskal-Wallis test is used to compare the distributions of SCORE for each borough.
• The hypotheses are as follows:
  • \(H_0\): The distributions of SCORE for borough \(i\) are equal.
  • \(H_1\): The distributions of SCORE for borough \(i\) are not equal.

KruskalResult(statistic=63.67589952287163, pvalue=4.890239593852517e-13)

The null hypothesis is rejected at \(\alpha = 0.01\).

• Dunn’s Test is a post-hoc test which performs multiple pairwise comparisons between the distributions of SCORE for each borough.
• The hypotheses are as follows:
  • \(H_0\): The distributions of SCORE for borough \(i\) and borough \(j\) are the same.
  • \(H_1\): The distributions of SCORE for borough \(i\) and borough \(j\) are different.

• Random Forest Regression is used to predict the inspection score.
• Preprocessing steps include:
  • Dummy variables are created for categorical variables.
  • ‘GRADE’ is converted to a numeric variable using ordinal encoding.
  • The data is split into training and testing sets.

• Average Baseline error is calculated from the mean of the absolute difference between the average training score and the testing score.

Average Baseline Error: 4.26

Models are created using RandomForestRegressor from sklearn.ensemble

The first model rf is created using \(n = 1000\) trees.

The parameters of the model are as follows:

{'bootstrap': True, 'ccp_alpha': 0.0, 'criterion': 'squared_error', 'max_depth': None, 'max_features': 1.0, 'max_leaf_nodes': None, 'max_samples': None, 'min_impurity_decrease': 0.0, 'min_samples_leaf': 1, 'min_samples_split': 2, 'min_weight_fraction_leaf': 0.0, 'n_estimators': 1000, 'n_jobs': None, 'oob_score': False, 'random_state': 19, 'verbose': 0, 'warm_start': False}

Results:

Mean Squared Error: 5.72
Mean Absolute Error: 1.26
Root Mean Squared Error: 2.39
R^2: 0.8846

The second model gs_rf is created using GridSearchCV to find the optimal parameters for the model.

The parameters of the model are as follows:

{'bootstrap': True, 'max_depth': 80, 'max_features': 3, 'min_samples_leaf': 3, 'min_samples_split': 12, 'n_estimators': 100}

Results:

Mean Squared Error: 11.50
Mean Absolute Error: 1.75
Root Mean Squared Error: 3.39
R^2: 0.7683

The third model rs_rf is created using RandomizedSearchCV to find the optimal parameters for the model.

The parameters of the model are as follows:

{'n_estimators': 1400, 'min_samples_split': 5, 'min_samples_leaf': 2, 'max_features': 'auto', 'max_depth': None, 'bootstrap': True}

Results:

Mean Squared Error: 5.55
Mean Absolute Error: 1.26
Root Mean Squared Error: 2.36
R^2: 0.8881

The Randomized Search model has the highest \(R^2\) value, and the lowest MSE and RMSE values. Along with the rf model, the rs_rf model has an MAE value that is substantially lower than the baseline error.

The permutation importance is given below: